Comparative Analysis of Takagi-Sugeno and Mamdani Fuzzy Inference Architectures with ANFIS-Based Automated Rule Generation for EEG-Based Cognitive State Monitoring

Abstract

Fuzzy inference systems have demonstrated considerable promise for EEG-based cognitive state monitoring in neurodegenerative conditions. However, two design decisions significantly influence system performance and clinical applicability: the choice of inference architecture (Takagi-Sugeno vs Mamdani) and the method of rule and membership function generation (manual expert-driven vs data-driven automated). This paper presents a comparative analysis of both dimensions in the context of an EEG-based Alzheimer’s disease monitoring system operating on the ds004504 OpenNeuro dataset (88 subjects: 36 AD, 23 FTD, 29 HC). A Takagi-Sugeno system, implemented as a hybrid FSM-Fuzzy architecture, is compared against a Mamdani equivalent across four axes: inference transparency, output interpretability, computational complexity, and suitability for FPGA deployment. The limitations of manual membership function tuning are then examined through failure case analysis, and an Adaptive Neuro-Fuzzy Inference System (ANFIS) proof-of-concept is presented, demonstrating automated rule and MF generation from the delta/alpha ratio (DAR) biomarker. ANFIS training on 88-subject DAR data (70 training, 18 test) achieved minimal training RMSE of 0.392 and validation RMSE of 0.391. Data-driven MF centers differ substantially from manually designed centers — shifts of 2.754σ, 0.455σ, and 2.055σ for LOW, MEDIUM, and HIGH MFs respectively — quantifying the representational gap that expert tuning introduces. These findings motivate a hybrid design strategy: ANFIS for offline MF calibration, TS for online FPGA inference, and Mamdani for clinical output presentation.